[Build] Add support for Ascend950 chip

[linfeng-yuan]

What this PR does / why we need it?

This PR adds support for the Ascend950 chip. This includes:

• Updating build scripts (CMakeLists.txt and setup.py) to recognize the Ascend950 chip and set appropriate compilation flags.
• Disabling a set of custom operators that are not yet supported on the Ascend950 hardware target.
• Performing a codebase-wide refactoring of pipe_barrier() calls to the namespaced AscendC::PipeBarrier<>() for improved code consistency and adherence to the latest API standards.

Does this PR introduce any user-facing change?

Yes, this PR adds support for a new hardware target, the Ascend950.

How was this patch tested?

Ascend950DT e2e passed (Qwen3-32B-MXFP8) and CI passed

• vLLM version: v0.16.0
• vLLM main: vllm-project/vllm@4034c3d

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request primarily extends hardware support to include the Ascend950 chip. This involves modifications to the build configuration to correctly identify the new chip and manage operator compatibility. Additionally, a widespread refactoring was performed to update the pipe_barrier synchronization primitive to use the more explicit AscendC::PipeBarrier API within custom kernel implementations, enhancing consistency and potentially future-proofing the code.

Highlights

• Ascend950 Chip Support: Introduced build system logic to recognize and configure for the Ascend950 chip, including specific operator exclusions and compile definitions in CMake.
• Pipe Barrier API Standardization: Refactored numerous custom operator kernels to standardize the usage of pipe barriers, replacing direct pipe_barrier calls with the more explicit AscendC::PipeBarrier API.
• Conditional Operator Compilation: Implemented conditional compilation for mla_preprocess and moe_grouped_matmul operators in the PyTorch binding, ensuring they are excluded when targeting Ascend950.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Changelog

• CMakeLists.txt
  • Renamed VLLM_ASCEND_CUSTOM_OP_EXCLUDE to VLLM_ASCEND_CUSTOM_OP_EXCLUDE_310P.
  • Introduced VLLM_ASCEND_CUSTOM_OP_EXCLUDE_ASCEND950 to exclude specific operators for Ascend950.
  • Added conditional logic to set _ASCEND950 compile definition when SOC_VERSION matches "ascend950".
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_fullload_dynamic_quant.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_fullload_quant.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_fullload_quant_base.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_gather_dynamic_quant.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_gather_quant.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_init_routing_fullload.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_mrgsort_out.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_sort_multi_core.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_sort_one_core.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_src_to_dst_and_gather.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/moe_init_routing_quant_v2/moe_v2_src_to_dst_op.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_ffn_combine/op_kernel/utils/hccl_shmem.hpp
  • Replaced pipe_barrier(PIPE_ALL) with AscendC::PipeBarrier<PIPE_ALL>().
• csrc/dispatch_gmm_combine_decode/op_kernel/dispatch_gmm_combine_decode/raw_distributed/cam_moe_distribute_combine.h
  • Replaced pipe_barrier(PIPE_ALL) with AscendC::PipeBarrier<PIPE_ALL>().
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/dispatch_gmm_combine_decode/op_kernel/dispatch_gmm_combine_decode/raw_distributed/cam_moe_distribute_dispatch.h
  • Replaced pipe_barrier(PIPE_ALL) with AscendC::PipeBarrier<PIPE_ALL>().
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/kernels/bgmv_expand.cpp
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/kernels/bgmv_shrink.cpp
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/kernels/sgmv_expand.cpp
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/kernels/sgmv_shrink.cpp
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/notify_dispatch/op_kernel/notify_dispatch.h
  • Replaced pipe_barrier(PIPE_ALL) with AscendC::PipeBarrier<PIPE_ALL>().
• csrc/sparse_flash_attention/op_kernel/sparse_flash_attention_service_vector_mla.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• csrc/torch_binding.cpp
  • Added conditional compilation (#ifndef _ASCEND950 and #endif) around mla_preprocess and moe_grouped_matmul operator definitions.
• csrc/utils/inc/kernel/pse.h
  • Replaced pipe_barrier(PIPE_V) with AscendC::PipeBarrier<PIPE_V>().
• setup.py
  • Updated get_chip_type function to correctly identify "950" in chip_name.
  • Modified gen_build_info to set device_type to "A5" if soc_version contains "ascend950".

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces support for the Ascend950 chip and performs a large-scale refactoring of pipe_barrier calls.

The key changes are:

• Ascend950 Support: Build scripts (CMakeLists.txt, setup.py) are updated to recognize the Ascend950 chip, setting the device type to 'A5' and adding a _ASCEND950 compile definition.
• Operator Disabling: A number of operators are disabled for the Ascend950 target by using the _ASCEND950 preprocessor directive in csrc/torch_binding.cpp. This is consistent with the build script changes that exclude the corresponding kernels from compilation.
• Refactoring: All occurrences of pipe_barrier() have been refactored to AscendC::PipeBarrier<>() across numerous kernel files. This appears to be a modernization effort to use a namespaced and templated version of the API, which improves code clarity and consistency.

The changes are logical and well-executed. The new chip support is implemented cleanly, and the refactoring is applied consistently.

As per the repository's style guide, I'm suggesting an updated Pull Request title and summary:

Suggested PR Title:

[Build][Feature] Add support for Ascend950 chip

Suggested PR Summary:

### What this PR does / why we need it?

This PR adds support for the Ascend950 chip. This includes:
- Updating build scripts (`CMakeLists.txt` and `setup.py`) to recognize the Ascend950 chip and set appropriate compilation flags.
- Disabling a set of custom operators that are not yet supported on the Ascend950 hardware target.
- Performing a codebase-wide refactoring of `pipe_barrier()` calls to the namespaced `AscendC::PipeBarrier<>()` for improved code consistency and adherence to the latest API standards.

### Does this PR introduce _any_ user-facing change?

Yes, this PR adds support for a new hardware target, the Ascend950.

### How was this patch tested?

CI should be run to ensure all existing tests pass on supported platforms. Specific tests should be added and executed on an Ascend950 device to validate the new support and ensure that the enabled operators function correctly.

[linfeng-yuan]

Currently, execution of some ops do not check enable_custom_op() and support fallback calling with torch_npu (e.g., init_routing_custom, gating_top_k_custom, etc.). Currently, we can use the pta API with CANN 8.5.1. We plan to investigate the affects and delete these unnecessary custom ops ASAP.

[linfeng-yuan]

OP_LIST: [enabled, disabled] add comments here

[linfeng-yuan]

We've added comments to refer reader to #7157 now.